Research On The Method Of Gait Feature Extraction And Recognition

The human identification based on the biometrics is currently one of the most active research topics in computer vision. Traditionally, techniques such as face, iris, fingerprint, DNA and handwriting recognition have already been utilized in numerous commercial and law enforcement applications. Gait, regarded as a new biometrics, has been recently used to recognize a person via his style of walking, which contains his physiological or behavioral characteristics. A unique .advantage of gait is the ability to operate at a distance, when other biometrics are of too low a resolution to be perceived. Furthermore, gait measurements are also non-intrusive and difficult to disguise and conceal in application scenarios (e.g. face can be obscured by helmets and fingerprints can be hidden by gloves). As a result, gait recognition is a potential solution for the applications of human identification, and a challenge in the area of computer vision.On the basis of its importance both in theoretical research and practical application, gait recognition needs to be further studied. With the help of current research findings, this thesis has made the following innovative contributions:1. Based on the perceptual shape descriptors and the eigenspace transformation, a new automatic gait recognition algorithm is proposed. The principal component analysis (PCA) and the multiple discriminant analysis (MDA) are used to reduce data dimensionality and to optimize the separability of different classes. The proposed algorithm is tested both on the UCSD database and the CMU databaSe. The average recognition rates of over 94% on the UCSD and 90% on the CMU databases are achieved respectively. Noticeably the result on the UCSD database is about 5% higher than that of the related literature [12].2. Oriented the region area feature and the region variance feature, a new gait recognition algorithm based on the region feature is proposed. The average recognition rate on the UCSD database is over 90% using the region area feature, and over 94% using the region variance feature, and about 1% and 5% higher than that of the Baseline algorithm respectively. This algorithm is also tested on the CMU database. The average recognition rate is 99.5% and 98.5% respectively, and about 4% and 3% higher than that of the Baseline algorithm.3. Oriented the dynamic variance feature and the dynamic energy feature, a new gait recognition algorithm based on the dynamic feature is proposed. The average recognition rates on the UCSD database are both about 92% using these two dynamic features, and are both about 1.2% higher than that of the related literature [57]. The average recognition rates on the CMU database are over 95% using the dynamic variance feature and 97% using the dynamic energy feature respectively, and 0.9% higher using the dynamic energy feature than that of the related literature [57].4. A novel gait recognition algorithm via fusing shape and kinematics features is proposed. The algorithm is carried out using two fusion strategies: feature level fusion and decision level fusion (the Sum and Product rules). The experimental results show that the correct recognition rates of the proposed fusion algorithms are improved to some extent, and the increase is from 4.3% to 23.8%.